Lessons from Linus Lee

Linus Lee, a research engineer, programmer, and writer, is a prominent voice in the fields of artificial intelligence, human-computer interaction, and tools for thought. Currently at Thrive Capital and formerly with Notion's AI team, he is known for his deep thinking on how technology can augment human creativity and agency. [1][2] Lee has built over 100 personal projects, including his own programming language and many of the tools he uses daily. [2][3] His work and writings explore language, knowledge work, and the future of software interfaces. [2][3]

On Artificial Intelligence & Machine Learning

On the goal of AI: "I want to build interfaces that let the AI gesture us into a better future without infringing on our agency." [1]
AI as a "thought calculator": Lee borrows from Simon Willison's concept of AI as a "thought calculator," a tool to enhance human imagination and creativity rather than a mere simulacrum of humans. [1]
The nature of intelligence in AI: He suggests that the goal of many corporations building AI is not to replicate human intelligence in its entirety—including its "annoying" parts like daydreaming—but to build an "intelligence kind of engine." [1]
The importance of experimentation with AI: Lee advocates for a playful and experimental approach to using AI tools like ChatGPT, encouraging users to "bang away at this thing and try everything possible" to understand its limits and capabilities. [1]
AI's potential to augment, not just automate: He is a proponent of designing AI to augment human abilities and help us live fuller lives, rather than simply replacing human tasks. [1]
On the limitations of current models: Lee acknowledges that the prototypes he builds are often not "good enough to actually be useful" for precise edits, highlighting the ongoing research needed to improve the models. [4]
The value of understanding the "raw data": He emphasizes the importance of spending time with the raw data and failure cases of models to deeply understand the problems you're trying to solve. [4]
Future improvements in language models: Beyond general reasoning, Lee sees hallucination reduction, better instruction following, and cost efficiency as key areas for improvement in language models. [4]
AI and the future of knowledge work: With language models, you can treat knowledge as a "blob of information that the system quote unquote knows," allowing for just-in-time synthesis of information. [5]
Chat is not the only interface: Lee is a vocal advocate for moving beyond chat-based interfaces for interacting with LLMs, believing that other modalities can offer more control and nuance. [6]

On Interfaces & Tools for Thought

Instrumental vs. Engaged Interfaces: Lee categorizes interfaces into two types: "instrumental," where the user just wants a result, and "engaged," where the user desires mastery and complexity. He argues that the right choice depends on the user and the task. [2]
The ideal instrumental interface: "The ideal instrumental interface for any task or problem is a magic button that can one read the user's mind perfectly to understand the desired task, and two perform it instantly, and completely to desired specifications." [2]
The danger of overly simplistic interfaces: Hiding complexity to make a tool easier to use can take away agency from the user. Lee cautions that doing this unintentionally is "probably bad." [7]
Interfaces for direct manipulation: He is interested in exploring ways to use language models that are more like familiar user interfaces such as "pinch-to-zoom or drag-and-drop" rather than just prompts. [8]
Representing concepts visually: Lee envisions a future where you can "see a paragraph and immediately grock oh that's about X, and the style is Y" through visual cues like color and shape. [6]
Bringing thought outside the head: The purpose of productivity and writing tools is to "represent concepts in a form that can be seen with the senses and manipulated with the body directly." [6]
Building your own tools for understanding: Lee is a firm believer in building your own tools to gain a deeper understanding of the problems you're trying to solve. [4][9]
The power dynamic in software: "There's kind of a power dynamic where if you don't know what the company that's providing you some software products is doing with your data, they have the power, whereas if you build your own thing, you understand exactly what's going on, you're in control." [9]
Software as a material: "Software is the closest thing you have to cheap, free material to build new things." [10]
The future of text interaction: His focus is on how we can improve the way people interact with text, moving beyond traditional reading and writing methods. [6]

On Learning & Knowledge

Learning by necessity: Lee's approach to learning new technologies is to "just kind of picking them up as I really really need it." [5][11]
Practical application over pure theory: When learning about deep learning, he found it more effective to have a paper open on one side of the screen and the PyTorch implementation on the other, even adding print statements to understand the code's execution. [11]
Moving past the capability walls of traditional techniques: He started with simpler NLP techniques but eventually hit their limitations, which pushed him to learn more complex deep learning methods. [5][11]
The difficulty of academic papers without context: He admits to having a hard time reading academic papers initially because he lacked the necessary vocabulary and background. [4]
Knowledge tools to aid learning: He has a side hobby and interest in building knowledge tools to help people learn and read quickly. [8]
The limitations of double-bracketing: He built an app that automatically highlights and links related ideas in notes as a response to the manual effort of double-bracketing in other note-taking systems. [8]
The importance of owning a niche: Lee advises to "research something deeply and own a niche," as this allows you to benefit from that knowledge indefinitely. [10]
Learning as a continuous process: His journey into AI has been self-directed and driven by a continuous desire to build and understand. [5]
The value of side projects for learning: His numerous side projects have been a primary vehicle for his learning and exploration in various domains. [5][10]
Expanding the domain of thought: The ultimate aim of his research is to "build interfaces and knowledge tools that expand the domain of thoughts we can think and qualia we can feel." [3]

On Work, Productivity & Creativity

Working on things you admire: "I try to work on things that I would admire myself for having worked on. If you do that, you usually end up in a pretty good place." [10]
The balance between work and reward: "There are things you do in life to earn the option of doing things you actually like... And if you keep working because it feels nice and never cash in, it's like saving a bunch of money even if you know you'll never use it. It's wasteful." [10]
The importance of ownership: "You can own something by having a bunch of side projects or writing as I do. You can also research something deeply and own a niche. A lot of people spend their time chasing validation and when you do that, you don't own anything." [10]
The trap of chasing validation: "You're selling yourself for the feeling of being interesting or important. But if you own something, you can benefit from it forever." [10]
The value of a prolific output: Lee has written over 400,000 words on his blog and worked on over 120 side projects, demonstrating a commitment to consistent creation. [10]
Context switching as a flow-breaker: He views fixing typos while writing as a form of context switching that can break your creative flow. [5]
The iterative nature of building: When building internal tools at Notion, they often built their own versions because by the time third-party solutions were mature, they had already developed their own systems. [5]
Moving quickly and experimenting: In the early phases of a company, the priority is to "move really quickly to do a lot of experiments to validate or invalidate what kinds of things work for people." [10]
The joy of creation: His passion for building things is a recurring theme, seeing software as an accessible medium for creation. [10]
Writing as a tool for thinking: His extensive writing on his blog serves as a way for him to process and develop his ideas. [3][12]

On Technology & The Future

Technology as an amplifier of agency: By default, technology tends to amplify the agency of those who already have resources, but it can be built in a more opinionated way to distribute that power. [2]
Building with intention: Lee urges technology builders to be thoughtful about the direction they are pushing the world, rather than "deriving your ideology as a function of what you end up sort of stumbling on into building." [2]
The future is not determined: "Technology is not determined: the future we imagine and create is entirely up to us. Will we optimize ourselves into something non-human, or dream our way into something beautiful?" [2]
The evolution of representations: He points to the shift from decimal to binary computing as an example of how simpler, more modular representations can win out over time. [6]
Beyond today's language and writing systems: Lee is interested in what comes after our current systems of language and how learning, creating, and collaborating will evolve in the long-term future of humanity. [3]
The potential for technology to make us more human: He hopes to create "instruments for super agency," leaning into the idea that technology at its best should make everyone more human and more capable. [2]
The seductive nature of "sexy" technology: He expresses concern that because "agents feel very sexy right now," problems that might be better solved through other means will be forced into an agent-based solution. [2]
The importance of diverse representations: Providing users with a diverse and accessible set of representations for what's happening within complex software systems can help them recover their agency. [7]
Technology's role in dreaming and wandering: He believes that technology can and should be used to help us dream and wander, not just optimize and automate. [2]
A thoughtful approach to progress: His work reflects a deep consideration for the humanistic implications of technology, advocating for a future where technology enhances our lives in meaningful ways. [1]

Learn more:

Lessons from Linus Lee

On Artificial Intelligence & Machine Learning

On Interfaces & Tools for Thought

On Learning & Knowledge

On Work, Productivity & Creativity

On Technology & The Future

Written by Antoine Buteau

Lessons from Mark Leonard

Lessons from Tom Siebel